Source-linked AI summary
Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems
Mengzhuo Chen, Junjie Wang, Fangwen Mu, Yawen Wang, Zhe Liu, Huanxiang Feng, Qing Wang
TL;DR
Failure attribution in LLM-based MASs is difficult because existing benchmarks often expose only outputs, omitting inputs and context used in developer-facing debugging. The paper introduces TraceElephant, a benchmark with full execution traces and reproducible environments, and evaluates attribution techniques across configurations. Complete traces improve attribution over output-only traces, while dynamic replay further enhances attribution capability.
Problem
Existing MAS attribution benchmarks provide partially observable output-only traces, whereas developer-facing debugging commonly has access to inputs, intermediate states, tool interactions, and environment states.
Method
TraceElephant provides annotated full execution traces from representative agentic systems together with reproducible execution environments for evaluating failure attribution.
Results
Complete traces achieve 65.9% agent-level and 30.3% step-level accuracy, improving over output-only traces by 22% and 76%, respectively.
Takeaways & Limitations
TraceElephant provides a practical foundation for failure attribution research and evaluation practices that mirror real-world debugging.
Takeaways & Limitations
The study evaluates traces from only three MASs, so some findings may not generalize to all existing or future systems.
Abstract
from arXiv · showhide
Failure attribution, i.e., identifying the responsible agent and decisive step of a failure, is particularly challenging in LLM-based multi-agent systems (MAS) due to their natural-language reasoning, nondeterministic outputs, and intricate interaction dynamics. A reliable benchmark is therefore essential to guide and evaluate attribution techniques. Yet existing benchmarks rely on partially observable traces that capture only agent outputs, omitting the inputs and context that developers actually use when debugging. We argue that failure attribution should be studied under full execution observability, aligning with real-world developer-facing scenarios where complete traces, rather than only outputs, are accessible for diagnosis. To this end, we introduce TraceElephant, a benchmark designed for failure attribution with full execution traces and reproducible environments. We then systematically evaluate failure attribution techniques across various configurations. Specifically, full traces improve attribution accuracy by up to 76\% over a partial-observation counterpart, confirming that missing inputs obscure many failure causes. TraceElephant provides a foundation for follow-up failure attribution research, promoting evaluation practices that reflect real-world debugging and supporting the development of more transparent MASs.
1 Introduction
LLM-based MASs complicate failure attribution through natural-language, nondeterministic interactions, while existing output-only benchmarks omit developer-relevant execution context. TraceElephant addresses this gap with full traces, reproducible environments, and systematic evaluation of attribution techniques.
- Motivation: LLM-based MASs challenge traditional failure attribution through interactions among agents, tools, and internal reasoning processes.Traditional techniques assume discrete states, traceable executions, and largely deterministic component behavior.
- Motivation: Existing Who&When traces expose agent outputs but omit inputs such as task instructions, prompts, and contextual messages.This setting suits some black-box scenarios but differs from developer-facing debugging, where richer execution information is typically available.
- Benchmark: TraceElephant provides step-by-step traces containing actions, natural-language inputs and outputs, tool interactions, configurations, and architecture, alongside reproducible execution environments.The environment supports controlled re-execution, state inspection, and interactive hypothetical debugging queries.
- Evaluation: 65.9% agent-level and 30.3% step-level attribution accuracy with complete traces improved by 22% and 76%, respectively, over output-only traces.Running environments further improved step-level accuracy by 10%.
- Implications: The results support using full execution details for attribution and developing more diverse benchmarks for cumulative evaluation of MAS failure attribution.Performance also varies across architectures, agent types, and step positions.
- Benchmark: 220 failure traces from three representative agentic systems are annotated with the responsible component and decisive failure step.The systems include multi-agent orchestration and a single-agent tool-centric scaffold.
2 Problem Definition
The paper models an MAS as functional components acting sequentially under a turn-based protocol. Failure attribution identifies both the earliest unrecoverable failure step and the responsible component, using recoverability and component roles rather than chronology alone.
- System model: An MAS is represented as a finite set of functional components collaboratively performing a task through discrete, turn-based steps.Exactly one component acts at each step, and components may be explicit agents or modules in a single-agent scaffold.
- Attribution target: Step-level attribution identifies the earliest point where failure becomes inevitable, while agent-level attribution identifies the responsible component at that step.The definition is role-aware and recoverability-aware rather than purely chronological.
- Formalization: The formalization and notation for the attribution problem are provided in Appendix A.1.The main text states the goal in terms of step-level and agent-level responsibility given the full failure trace.
3 Benchmark Construction
TraceElephant is built from executable agentic systems and fully observable traces paired with annotated responsibility labels. Its collection pipeline preserves execution fidelity while recording trace metadata, step inputs and outputs, tool interactions, and expert-validated annotations.
- Dataset scope: Each benchmark instance pairs an executable MAS and fully observable execution trace with the responsible agent and decisive failure step.The benchmark applies across dynamically assembled teams, fixed-role orchestration, and a single-agent tool-centric scaffold.
- Data sources: TraceElephant collects traces from three representative agentic systems spanning dynamic teams, centralized orchestration, and a single-agent tool-centric scaffold.Task designs are aligned with each system’s intended capabilities.
- Collection pipeline: A lightweight middleware captures LLM requests, responses, and tool interactions without modifying the original agent implementations.Pre-processing extracts basic attributes while preserving the original execution flow.
- Trace contents: Each instance includes the execution trace, runnable code, and configurations for the specific trace.Trace-level metadata records the task, system, agent configuration, and system architecture.
- Trace contents: Step-level records store ordered agent actions with input context, outputs, and tool logs containing tool names, arguments, outputs, and execution status.The trace structure separates input and output fields for each step.
- Annotation: Expert annotation labels the responsible component and failure-origin step, achieving Krippendorff’s alpha of 0.72 for agent labels and 0.64 for step labels before consensus review.Uncertain cases are jointly reviewed and finalized through expert consensus.
4 Failure Attribution Evaluation
TraceElephant evaluation shows that attribution performance depends on observability, configuration, backbone model, architecture, and failure position. Full traces and dynamic verification particularly improve fine-grained step attribution, while system-specific patterns motivate architecture-aware methods.
- Static vs. Dynamic Attribution Performance: Dynamic configuration reaches 33.3% step-level and 66.7% agent-level accuracy, compared with 30.3% and 65.9% for static configuration.Dynamic verification improves step-level attribution by 10% through controlled re-execution and counterfactual probing, while agent-level gains remain limited.
- Ablation Study: 76% versus 22% accuracy drop shows that missing information harms step-level attribution more than agent-level attribution.Removing metadata or input fields causes noticeable degradation, reflecting the greater context dependence of identifying the decisive failure step.
- Effect of Backbone LLMs: Claude-4.5-Sonnet, DeepSeek-R1, and GPT-4o perform relatively strongly, whereas Qwen3-32B and GPT-OSS-20B perform more weakly.The paper associates the contrast with differences in reasoning architecture, contextual understanding, parameter scale, and sustained multi-step reasoning capacity.
- Failure and Attribution Patterns: External-environment and concrete-operation agents account for almost over 50% of failures, while failure distributions differ across system architectures and execution phases.Automatically generated teams show dispersed failures, whereas manually crafted systems concentrate failures early; Magentic-One step accuracy rises from 8% early to 52% late.
- Agent-level Accuracy: Data-processing agents reach 53% accuracy in Captain-Agent, web-related agents 22%, and the Magentic-One Orchestrator 38%.Tool-mediated traces make some failures more discernible, while ambiguous natural-language search outputs make web-agent attribution harder.
- TraceElephant vs. Who&When: Output-only traces reduce accuracy from 62% to 51% at the agent level and from 28% to 16% at the step level.The output-only setting corresponds to the partial-observation regime represented by Who&When.
- Implications and Takeaways: Architecture-aware attribution is motivated by variation in responsible agent types, decisive-step positions, and attribution performance across MAS designs.The paper highlights centralized versus dynamic team formation and tool-heavy versus planning-heavy workflows as relevant design distinctions.
- Future Directions: Static Agentic attribution currently uses basic tool inspection, while dynamic evaluation primarily performs single-step replay and counterfactual checks.The paper identifies richer hypothesis testing, graph reasoning, control-flow reconstruction, and state-space exploration as directions for deeper analysis.
5 Related Work
Related work includes attribution techniques for MASs and benchmarks intended to evaluate them. TraceElephant is positioned against limited existing benchmark coverage and the need for realistic evaluation resources.
- LLM-based Multi-Agent Systems: LLM-based MASs solve complex tasks through natural-language interaction, planning modules, and tool usage across architectures such as centralized, decentralized, and hybrid coordination.These systems are studied for software engineering, question answering, and decision-making tasks.
- Failure Attribution in MAS: Existing MAS attribution techniques include spectrum analysis, reinforcement learning, and graph-based modeling of agent interactions.The cited examples are FAMAS, AgentTracer, and GraphTracer.
- Benchmarking: High-quality benchmarks are presented as essential for evaluating existing attribution techniques and supporting new ones.The benchmark motivation is tied to reliable evaluation of failure attribution in MASs.
6 Conclusion
TraceElephant addresses failure attribution in LLM-based MASs under full execution observability. The authors report improved attribution performance and position the benchmark as a foundation for research that reflects real-world debugging.
- Full execution observability improves attribution performance in LLM-based MASs.The conclusion also reports that dynamic replay further enhances attribution capability.
- TraceElephant provides a practical foundation for future failure attribution research.
- The benchmark promotes evaluation practices that mirror real-world debugging.
7 Limitations
The study evaluates developer-facing failure attribution using traces from three representative multi-agent systems. Its findings may not generalize to all existing or future systems or black-box usage scenarios.
- The study focuses on developer-facing attribution with execution traces from a limited set of representative multi-agent systems.It does not cover all possible system architectures or black-box usage scenarios.
- Findings may not generalize to all existing or future systems because the study evaluates only three MASs.The systems are Captain-Agent, Magentic-One, and SWE-Agent.
- The selected systems span dynamic team assembly, centralized orchestration, and specialized software engineering workflows.This diversity is intended to enhance benchmark representativeness and mitigate architecture-specific bias.
A.1 Detailed Problem Definition
The formalization models a MAS as a discrete-step execution that produces a structured trace and task outcome. Failure attribution identifies the earliest inevitable failure step and the responsible agent at that step.
- Multi-Agent Execution and Observability: At each discrete step, a selected agent receives an input, produces an output, and updates the system context.The execution terminates after T steps and produces a trace of execution steps.
- Multi-Agent Execution and Observability: Each execution step is represented by an observable record containing the input, output, acting agent, step identifier, and agent identifier.
- Task Outcome and Failure Attribution: A task outcome is successful when the completed trace satisfies the task and failing otherwise.The outcome function maps the completed trace to 1 for success and 0 for failure.
- Task Outcome and Failure Attribution: Step-level attribution identifies the earliest step after which every feasible continuation leads to failure.This is an inevitability criterion over continuations of the trace prefix.
- Task Outcome and Failure Attribution: Agent-level attribution identifies the agent acting at the failure step as responsible for the failure.
A.2 Details about Data Sources
TraceElephant draws traces from three diverse agentic systems across general-purpose reasoning and software engineering tasks. The dataset uses repeated fixed-configuration runs and preserves naturally occurring failure distributions.
- Systems: The benchmark includes Captain-Agent, Magentic-One, and SWE-Agent.These represent dynamic team assembly, fixed-agent orchestration, and a single-agent tool-centric software engineering scaffold.
- Task Sources: Captain-Agent and Magentic-One use GAIA and AssistantBench tasks involving multi-step reasoning, information gathering, and tool use.
- Task Sources: SWE-Agent uses SWE-Bench Verified tasks focused on code navigation, modification, and validation.These tasks involve repeated interactions between language-model agents and external development environments.
- Task Sources: SWE-Bench failures commonly involve incorrect code-change localization, incomplete fixes, or mismatches with test requirements.
- Run Configuration: Each system-task pair uses fixed configuration, temperature 0.3, a unique run identifier, and multiple trials.
- Trace Statistics: 380 cleaned execution traces were collected, including 220 failed traces used as benchmark instances.The benchmark preserves naturally occurring failure distributions rather than enforcing a uniform failure rate.
A.3 Details about Trace Collection
TraceElephant collects heterogeneous MAS executions through a unified logging pipeline that preserves LLM requests, responses, agent identities, step order, and tool-mediated interactions. Lightweight preprocessing structures these traces without modifying the original systems.
- TraceElephant adapts different MAS implementations to a unified execution-logging interface without changing their original implementations.
- A transparent LLM API middleware intercepts agent requests, forwards them to a selected backend model, and logs the requests with their responses.This provides observability of LLM-mediated decision making while preserving agent logic, prompts, and control flow.
- Lightweight regular-expression preprocessing extracts agent names and categorizes outputs as plain LLM responses or tool-mediated interactions.
- Agent identities are extracted from execution-time markers, then linked consistently to chronological step identifiers within each trace.
- Tool-mediated steps are identified from invocation patterns in output text, such as an action, tool input, and external search request.
A.4 Example Execution Trace
The example trace shows how TraceElephant records a multi-agent travel-planning execution as ordered, fine-grained steps with inputs, outputs, metadata, and raw tool interactions. It also illustrates attribution labels, evaluation metrics, and benchmark comparisons associated with the trace data.
- The travel-planning execution proceeds through decomposition, information retrieval, plan drafting, feasibility verification, and final recommendation.PlannerAgent, SearchAgent, and VerifierAgent perform these stages across six recorded steps.
- SearchAgent steps preserve raw web-search inputs, outputs, tool names, and execution status alongside the agent’s textual output.
- TraceElephant’s example records task metadata, system identity, agent composition, architecture, prompts, tools, and ordered step-level execution records.The trace is stored in a unified JSON format containing rich metadata, fine-grained inputs and outputs, and raw tool interactions.
- Human annotators identify the failure-responsible agent and decisive failure step by inspecting complete traces, with uncertain cases jointly reviewed until consensus.A final cross-review checks consistency with the shared annotation standards.
- Attribution is evaluated with agent-level and step-level accuracy, including a tolerance-based step metric, with results averaged over three independent runs.
- TraceElephant’s traces contain more average LLM invocations than Who&When for CaptainAgent and slightly more for Magentic-One: 20.5 versus 9.6 and 29.3 versus 28.8.